Skip to main content
Back to timeline
arXivSource publication:

Two policies for one rendezvous: a learned mixture guidance law meets 10 m and 1 m/s on all 600 multi-revolution low-thrust test missions

Synopsis

Multi-revolution low-thrust rendezvous must combine a long orbital transfer with meter-level terminal control, and a single policy struggles to cover both state scales. This guidance law has a far-range expert read modified equinoctial elements and a terminal expert read the relative state in the target's local frame, while a temporal-attention gate blends their thrust commands continuously. On 600 test missions it completed every rendezvous at 10 m and 1 m/s tolerances, cutting mean flight time by 50.6% and velocity increment by 44.5% relative to 4 km distance-based hard switching.

AI-generated editorial illustration: Temporal Attention-Enhanced Mixture-of-Experts Guidance for Multi-Revolution Low-Thrust Rendezvous

Interpretation

Multi-revolution low-thrust rendezvous can now be closed in loop by a guidance law trained offline whose onboard use is only a forward evaluation, instead of repeatedly solving an optimal-control problem in flight. Earlier routes either solved long-horizon optimal control online or switched by a distance threshold between a far-range and a terminal policy, with weights restricted to 0 or 1. On 600 test missions the jointly fine-tuned mixture-of-experts guidance succeeded in all 600 at 10 m and 1 m/s tolerances; on the 599 missions both it and 4 km hard switching completed, mean flight time fell by 50.6% and velocity increment by 44.5%.

The far-range and terminal phases each use the state description suited to their scale, and a learned gate allocates control authority continuously between them. Distance-based hard switching allows only a binary choice set by a range threshold, and fixed-weight blending applies the same weights throughout the mission, so neither lets both experts contribute over a transition region. With identical pretrained experts, the learned gate cut mean flight time by 26.4% and velocity increment by 16.9% relative to 4 km hard switching on the 599 commonly successful missions.

The terminal expert and the gate both read a window of recent observations, encoded by a Transformer, to supply target-orbit information absent from a single frame. A single frame holding the relative state and target mean motion does not determine target eccentricity or orbital position, so the same observation can correspond to different subsequent relative motion. On 600 terminal-phase initial conditions drawn from a held-out handover pool, the attention-based terminal expert succeeded in all 600, while the single-frame multilayer-perceptron version succeeded in 291 and the other 309 reached the 2000-interval limit.

Perspective

The results cover low-Earth-orbit non-coplanar elliptical rendezvous with arbitrary initial phase difference under a given thrust-acceleration bound, with success defined at 10 m and 1 m/s and a mission limit of about 92.6 days; the robustness tests apply navigation error, unmodeled perturbations, eclipse outages and 20% thrust loss each separately on top of the unmodeled-perturbation case, with one noise realization per mission.

Eclipse outages and thrust loss are added separately to the unmodeled-perturbation case, so their simultaneous occurrence and larger measurement errors are not assessed; navigation-filter dynamics and attitude and actuator errors remain for future work. In the illustrative mission, a fixed-final-time direct optimization initialized from the guidance trajectory lowers the velocity increment from 302.561 m/s to 252.380 m/s, indicating a remaining cost of feedback guidance relative to an offline, mission-specific solution.